Front-vehicle ranging method based on monocular vision and image segmentation under vehicle-borne camera

Abstract

The invention discloses a front-vehicle ranging method based on a monocular vision and image segmentation under a vehicle-borne camera. The method comprises a step of extracts a two-dimensional bounding box and a three-dimensional bounding box from a target vehicle based on a depth learning algorithm to obtain corresponding position information, a step of matching a 3D CAD vehicle model based on the length, width and height of the three-dimensional bounding box and obtaining a similar 3D vehicle model of a corresponding vehicle, a step of extracting vehicle model classification information ona vehicle in a picture based on the two-dimensional bounding box, a step of sending 3D information and vehicle model information corresponding to the vehicle to an example segmentation network, and calculating the absolute depth value of the vehicle in the image is calculated according to the size information of different vehicle models according to the camera imaging principle. According to the method, the time of the depth calculation is saved, a driving view is ensured, and so, a driver can intuitively observe a specific distance value of a front vehicle and make a reasonable driving judgment.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a distance measuring method for a preceding vehicle based on monocular vision and image segmentation under a vehicle-mounted camera. Background technique [0002] The popularity of automobiles has increased the world's demand for intelligent vehicles, and the development of machine vision has enabled machines to obtain image information in the same way as human eyes. [0003] Depth estimation based on video images is a prerequisite for realizing automatic driving and safe driving of vehicles. The so-called image depth estimation is a method of obtaining the actual distance of corresponding objects in the real world based on the information of the two-dimensional image stream obtained by the video equipment. Most of the traditional ranging methods need to calibrate the internal and external parameters of the camera and the camera height. The current ranging method based on...

AI Technical Summary

Problems solved by technology

This kind of method uses color to represent the relative depth. In terms of the need to obtain the integrity of the depth, it has certain advantages in the confined space range such as indoor distance measurement. However, for actual traffic scenes, redundant items and The correlation is weak in the process of traffic driving, and the advantage of detecting its depth information in the open space is not great, and it may even reduce the detection efficiency, and there is a small change in the contrast color when the color is used to indicate the distance, it is difficult to judge the specific distance of the vehicle

When the distance between the measured object and the camera is relatively close, the depth can be more clearly represented by the color, but for the distant object, due to the large range of color changes, it is difficult for the naked eye to identify the specific depth information of the distant object and when Vehicle information cannot be given completely when there are many vehicles or vehicles overlap

In real traffic scenarios, using color differences to express relative depth information cannot provide real security for users

[0005] In a real traffic scene, the traditional distance measurement method only uses the geometric projection relationship to estimate the image depth. This kind of method needs to calibrate different equipment and is easy to introduce errors. For the method of using the lidar instrument for distance measurement, due to the specific The detectable distance is limited in certain places and new equipment needs to be introduced. In terms of cost and convenience, this method is not the best method for current distance measurement.

The distance measurement method commonly used by neural networks uses the overall distance measurement of the image and uses the color difference to indicate the distance relationship between the vehicles in the image. It is not conducive to the observation of the vision of the road ahead

[0006] Looking at most of the existing depth estimation methods based on neural networks, there are two problems: 1) estimating relative depth rather than absolute depth; 2) depth information contains a large number of redundant items, such as the sky, distant buildings and two Sidewalk trees, etc.

Images